The 10-day AI/ML Corporate Training program led by Edoxi equipped the Tanzanian Government Team with the skills to develop and deploy a bilingual, on-premises AI helpdesk.
Government / public sector
Build & Deploy AI/ML Models on Web & Mobile Applications - 10-day corporate AI/ML programme
Tanzania, East Africa (delivered by Edoxi Training Institute, Dubai)
The program aimed to deliver instant, accurate answers from hundreds of policy, tender, and procedural documents, while building a fully in-house AI capability. Within 10 days, participants gained the skills to develop, own, and maintain the solution independently.
The project faced four key challenges: strict data sovereignty requirements, bilingual support for English and government Swahili, varying participant skill levels, and the need to run AI models efficiently without GPU hardware.
By Day 10, the team had deployed a live, bilingual on-premises AI helpdesk with source-cited responses for transparency and accuracy. The complete solution, including code, deployment assets, and documentation, was handed over, enabling the team to operate, maintain, and extend the system independently.
A government organisation in Tanzania ran on documents, procurement policies, tender papers, human resources manuals, terms and conditions, FAQ guides, public notices, and organisational profiles. When a member of staff or a member of the public needed a straightforward answer, someone had to find it first, usually by reading through long PDFs. The information existed; getting to it took time the team could not afford.
The organisation wanted to fix this with an AI assistant that could read its documents and answer questions directly. But it could not take the obvious route, and four hard constraints shaped every design decision that followed.
Data Sovereignty
Cloud AI was ruled out. All documents, queries, and responses had to remain within the government's internal network at all times.
Bilingual Service
Government work in Tanzania runs in both English and Swahili. The system had to serve either language with equal quality, not as a translation add-on.
Mixed Skill Cohort
Participants ranged from staff with no machine-learning background to intermediate developers, requiring a unified, graduated learning path.
No GPU Available
Training laptops lacked graphics processing units, making large model inference slow and requiring CPU-optimised model choices throughout.
Edoxi delivered a 10-day programme, “Build & Deploy AI/ML Models on Web & Mobile Applications,” built around a simple principle: participants would learn AI by building a real-world system from day one rather than studying concepts in isolation. Every topic introduced during the first six days was selected because it directly contributed to the solution developed in the final four days. Working from an authentic government brief, with real constraints around data availability, multilingual requirements, compliance, and infrastructure, ensured that participants encountered the same considerations and trade-offs faced in real AI projects.
The programme was structured in two phases. Days 1–6 focused on the foundations of artificial intelligence, covering machine learning, deep learning, natural language processing, and the transformer architectures that power modern large language models (LLMs), using public-sector and banking examples to maintain relevance. Days 7–10 shifted from theory to execution, with participants applying their knowledge to design, build, and deploy a working AI solution, integrating models into web and mobile applications while addressing the practical challenges of a real-world implementation.
Participants wrote production-grade code, deployed real containers, and demonstrated a fully functioning system by the end of the programme. To accommodate varying levels of technical experience, each day followed a consistent structure: introducing concepts, building code collaboratively, and integrating that code into the larger solution. This approach enabled beginners to follow a clear learning path while allowing more experienced participants to extend and enhance the implementation. By working within a shared codebase, the cohort collaborated on a single end-to-end system rather than a collection of isolated exercises.
The solution developed, Helpdesk-AI, is an on-premises intelligent assistant that enables users to query organisational PDF documents in English or Swahili. Built using a Retrieval-Augmented Generation (RAG) architecture, the system first retrieves the most relevant content from the document repository before generating a response, with every answer linked to its source document and page. This design improves accuracy, transparency, and trustworthiness by ensuring responses remain grounded in verified organisational information.
The hardest constraint, data sovereignty, was solved by building the system to be provider-agnostic. During training, the team used a commercial AI service for quick demonstrations, at a total cost of roughly three to five US dollars for the whole course. For production, the same code switches to a model running entirely on the government's own server through a single configuration change. No rewrite, no second codebase.
Swahili was the next problem. Standard models handle everyday Swahili reasonably but struggle with formal government language. The team used a multilingual embedding model for on-premises use, then applied Low-Rank Adaptation (LoRA) to a small set of government question-and-answer examples. Thus, the model picked up the right tone and terminology.
By the final day, the team had containerised the whole system with Docker so it deploys with a single command, added an automated workflow that classifies and routes incoming complaints, and built audit logging that records every upload, deletion, and query for public-sector accountability. Edoxi handed over the full package: source code, deployment scripts, a maintenance guide, and a security checklist.
The programme concluded with a post-training assessment to evaluate participants’ understanding of key AI and ML concepts and their ability to apply them in practice. Feedback was collected to assess the effectiveness of the training and identify areas for improvement. Upon successful completion, participants received certificates recognising their achievement and practical contribution to the development of the Helpdesk-AI system.
The government team did not watch the trainer build a system. They built it themselves, on their own infrastructure, under their own data rules, in the two languages their staff actually use. They finished the ten days with a working helpdesk and, just as important, with the source code and the understanding to run it, update it, and extend it without calling anyone. The capability stayed inside the organisation.
